© 2013, American Marketing Association Journal of Marketing
PrePrint, Unedited
All rights reserved. Cannot be reprinted without the express permission of the American Marketing Association.
THE EFFECTS OF POSITIVE AND NEGATIVE ONLINE CUSTOMER REVIEWS: DO BRAND STRENGTH AND CATEGORY MATURITY MATTER?
Nga N. Ho-Dac, Stephen J. Carson, and William L. Moore
Nga N. Ho-Dac ([email protected]) is Assistant Professor of Marketing, School of Business, Dalton State College,650 College Drive, Dalton, GA 30720, (801) 915-6864. Stephen J. Carson ([email protected]) is Associate Professor and David Eccles Faculty Fellow and William L. Moore ([email protected]) is the David Eccles Professor of Marketing, David Eccles School of Business, University of Utah, 1655 E. Campus Drive, Salt Lake City, Utah 84112. The authors would like to thank the members of the
reviewing team for their valuable suggestions. Please address all correspondence to the first author.
2
THE EFFECTS OF POSITIVE AND NEGATIVE ONLINE CUSTOMER REVIEWS: DO BRAND STRENGTH AND CATEGORY MATURITY MATTER?
Abstract
Brand equity is shown to moderate the relationship between online customer reviews (OCRs) and sales in both the emerging Blu-ray and mature DVD player categories. Positive (negative) OCRs increase (decrease) the sales of models of weak brands (i.e., brands without significant positive brand equity). In contrast, OCRs have no significant impact on the sales of the models of strong brands, although these models do receive a significant sales boost from their greater brand equity. Higher sales lead to a larger number of positive OCRs, and increased positive OCRs aid the transition from a weak to a strong brand. This creates a positive feedback loop between sales and positive OCRs for models of weak brands that not only helps their sales but also increases overall brand equity, benefitting all models of the brand. Contrary to the view that brands matter less in the presence of OCRs, we find that OCRs matter less in the presence of strong brands. Positive OCRs function differently than marketing communications in that their effect is greater for weak brands.
Keywords: Online customer reviews (OCRs), user-generated content (UGC), brand equity, category maturity, word-of-mouth (WOM).
3
Easy access to online customer reviews has led some observers to suggest that alternative assurances of product quality and performance such as brands will lose much of their importance in the interactive marketing environment (e.g., Chen 2001). This line of reasoning suggests that customers will bypass marketer influenced signals like brands and instead rely directly on unfiltered word-of-mouth from other consumers. Since the information contained in OCRs does not originate with the company, it is generally considered to be highly credible and influential (Bickart and Schindler 2001). Therefore, it is possible that this long-tail perspective will hold and consumers will use OCRs to find desired products irrespective of their brand name. On the other hand, the marketing literature offers evidence on the importance of brand equity which makes it improbable that brands will lose their value simply because consumers have access to OCRs.
In this research, we investigate the effects of brand equity and OCRs on sales response in an online selling environment. Of particular interest is how brand equity moderates the
relationship between OCRs and sales; i.e., whether OCRs affect the models of strong 1 or weak brands more. This issue is not as straightforward as it might appear since different literatures suggest different relationships. The brand equity and marketing communications literatures find evidence that strong brands have higher advertising elasticities and marketing communications effectiveness and are more protected from negative information (e.g., Ahluwalia, Burnkrant, and Unnava 2000; Belch 1981; Dawar and Pillutla 2000; Hoeffler and Keller 2003; Petty and
Krosnick 1995; Srivastava and Shocker 1991). If consumers respond to OCRs like advertising, strong brands should benefit more from positive OCRs and be hurt less by negative OCRs.
However, OCRs differ from marketer-sponsored communications in that they are more
4
credible (Cheong and Morrison 2008; Hung and Li 2007). Credibility suggests an active form of processing in which consumers evaluate the reliability of the source and its independence from the interests of the marketer. Viewed through the lens of signaling theory, the key issue is how the signal provided by OCRs compares to that provided by the brand. The brand signaling literature (Erdem and Swait 1998; Montgomery and Wernerfelt 1992) suggests that both positive and negative OCRs will affect weak brands more; positive OCRs provide a degree of credibility that weak brands cannot engender through company sponsored communications and negative reviews are evaluated without a compensating signal provided by a strong brand.
We also investigate how the effects of OCRs on strong and weak brands change across emerging and mature product categories. On one hand, mature categories feature large
cumulative numbers of OCRs which reduce uncertainty and increase the credibility of the information they contain. Hence, it is possible that strong brands will be less resistant to the influence of OCRs in mature categories. On the other hand, as categories mature, consumers learn more about the performance of brands within the category. Since brands store such
information, the category-specific equity of successful brands will tend to be stronger in mature categories, adding to the resilience of strong brands relative to OCRs. To examine this issue, we estimate models in both the emerging Blu-ray and mature DVD player categories. With the exception of maturity (at the time of data collection), these categories are very similar in terms of the mix of strong and weak brands, the number of models, price, etc.
The emerging literature on online word-of-mouth does not provide any direct evidence about how brand equity moderates the effect of OCRs or how this relationship changes over time. Much of the literature has used categories such as books, music, video games, and movies in which many products do not have pre-existing brand equity (e.g., Chen, Wu, and Yoon 2004;
5
Chevalier and Mayzlin 2006; Dellarocas, Zhang, and Awad 2007; Karniouchina 2011; Li and Hitt 2008; Liu 2006). These studies also provide similar recommendations for all products in a category without discrimination.
An exception to the latter is Zhu and Zhang (2010), who examine the interaction of OCRs and popularity on the sales of individual video games. They use two definitions of popularity. First, due to the typical precipitous sales decline over time, a game is defined to be popular if it has been on the market no more than four months. Second, a popular game is one with higher sales than the mean of all games in a given month. Under both definitions, they find that less popular games benefit more from OCRs than popular games.
In contrast, we examine the interaction between OCRs and brandequity, which we operationalize as the impact on sales of models of a brand that cannot be explained by other factors such as advertising, price, OCRs, competition, merchant, or model-specific effects. Brand equity differs fundamentally from the popularity of an individual model in that it is defined at the level of the product line. This allows us to consider the impact of OCRs not just on a specific model, but on all models in the product line as reviews for one model spill over to the brand
itself. In addition, while popularity can be influenced by other variables such as price and promotions in the short term, brand equity is defined after accounting for such factors and is therefore relatively robust. This makes it somewhat easier for managers to gauge with foresight when planning, particularly early in the product lifecycle.
Importantly, we expect our focus on brand equity to generate novel substantive
implications for managers operating in contexts where pre-existing brand equity is relevant—for example when managing product line extensions within a category, brand extensions from closely related product categories, or model updates. According to the perspective developed in
6
this research, even newly introduced models and models with high current or expected sales can be affected by OCRs, as long as consumers require reassurance because of a weak existing brand. In contrast, models of strong brands are affected less, even if they are poor individual sellers.
The results of the study indicate that brand equity moderates the relationship between OCRs and sales in both emerging and mature categories. Positive (negative) OCRs increase (decrease) the sales of models of weak brands but do not have a significant effect on the sales of models of strong brands. However, these models do receive a significant sales boost from being part of a strong brand. This is important, since positive OCRs for either all models or the leading model help build the equity of weak brands. Combined with the finding that greater sales lead to a larger number of positive, but not negative OCRs, this creates a positive feedback loop between
sales and positive OCRs for models of weak brands. Thus, positive OCRs help models of weak brands penetrate the market while simultaneously increasing the equity of the brand. This loop does not exist for the models of already strong brands, because they do not benefit to the same degree from positive reviews.
More broadly, in contrast to the view that brands matter less in the presence of OCRs, we find that OCRs actually matter less for strong brands. OCRs increase the sales of models of weak brands, and help weak brands become strong brands, but they do not affect brands to the same degree once they become strong. The sales boost that models receive from being part of a strong brand further indicates that brand equity is extremely important, even when the effect of model-level OCRs is controlled. In addition, the findings show that positive OCRs function differently than marketing communications in that their effect is greater for weak brands than strong brands. Finally, we find weaker effects of negative reviews than positive reviews, both for the focal and
7
competing models, and that the number of both positive and negative OCRs that customers post first increases and then decreases over time.
As anticipated, the results lead to a very different set of managerial implications from Zhu and Zhang (2010). Whereas their results suggest that OCRs are more beneficial for
unpopular niche games and for games whose sales have fallen significantly from their peak, ours suggest a clear role for OCRs early in the product lifecycle for all models of weak brands. Managers of weak brands should focus on generating positive OCRs both for their direct effect on sales of the model as well as their indirect effect on building brand equity. These paths give weak brands a way to compete other than through traditional marketing communications.
In contrast to weak brands, additional positive OCRs do not further benefit the models of strong brands. Therefore, managers of these brands should not necessarily follow the same strategy used for weaker brands, or do so to the same degree. Instead, strong brands should pursue actions to build brand equity more directly (e.g., through advertising) rather than focusing too narrowly on OCRs. The results of Zhu and Zhang (2010) are more appropriate for movies, video games, books, etc. where there is not a strong pre-existing brand component and sales often peak at the introduction and decline from there. On the other hand, our results are more appropriate for products in a branded product line that have more traditional sales trajectories.
The remainder of the paper is organized as follows. In the next section we explore arguments in the brand signaling literature to hypothesize interactions between brand strength and OCRs in predicting sales response. Then, we describe the data and the empirical model. This is followed by results and a discussion of implications for theory and practice.
HYPOTHESES
8
performance creates risk (Erdem, Swait, and Valenzuela 2006). To cope with this risk, customers rely on signals to indicate product quality and performance when purchasing (Shimp and
Bearden 1982). Many marketing mix elements such as price (Stiglitz 1989; Tellis and Wernerfelt 1987), advertising (Kirmani 1990; Nelson 1974), and warranty (Boulding and Kirmani 1993) have been shown to serve as credible signals. Brands have been shown to be especially strong and effective signals of product quality (Erdem and Swait 1998; Rao, Qu, Ruekert 1999).
When confronted with reviews written by other users, it is likely that consumers will find such reviews en masse to be a highly credible source of information on product quality and
performance. While OCRs are believed to be more credible than marketing communications (Cheong and Morrison 2008; Hung and Li 2007), it is not clear whether they are more or less credible than brand equity, since reviews are written by individuals with incomplete information and varying motivations. What is known is that stronger brands provide more credible signals than weaker brands because they are more susceptible to the loss of established brand equity (Erdem and Swait 1998) and future sales and profit (Wernerfelt 1988). Thus, the OCR signal will tend to overshadow the limited brand signal for weak brands whereas for strong brands both signals will provide a degree of credible information. As a result, positive OCRs should have a larger effect on weak brands, which lack a credible brand signal, than strong brands, which already provide substantial assurance. Positive OCRs create a degree of credibility that weaker brands cannot create on their own.
In addition, signals like brand equity are important in decision making under uncertainty; i.e., in the absence of concrete evidence about product quality (Montgomery and Wernerfelt 1992). While positive reviews for models of strong brands largely reinforce consumer beliefs about these models and do little to reduce uncertainty, positive reviews for models of weak
9
brands not only help these models directly but also reduce the level of uncertainty about them. This tends to decrease the overall level of uncertainty facing consumers in the category, reducing the effect of strong brands and further benefiting weaker brands.
Finally, one of the arguments for why marketing communications are more effective for stronger brands is that the brand lends a degree of credibility to the advertisement. As noted above, extending this view to OCRs would suggest that positive OCRs would be more effective for strong brands. However, the credibility lent by a strong brand is less necessary for OCRs due to the inherent credibility of the reviews. Hence, the advantage associated with strong brands with respect to marketing communications is unlikely in the context of OCRs. Jointly, these arguments suggest that positive OCRs should benefit weak brands more than strong brands.
Hypothesis 1: Positive OCRs will have a stronger (positive) effect on the products of weak brands than those of strong brands.
Negative OCRs should also affect weak brands more than strong brands. As observed above, the brand equity and marketing communications literatures find that strong brands are more protected from negative information (Ahluwalia, Burnkrant, and Unnava 2000; Dawar and Pillutla 2000; Petty and Krosnick 1995; Srivastava and Shocker 1991). Similarly, from a
signaling perspective, strong brands possess an offsetting signal that is also highly credible to help overcome and buffer negative reviews. Weak brands, in contrast, lack a compensating signal; thus they will be affected to a greater extent.
In the case of negative reviews, the signaling and branding/communications perspectives lead to the same prediction. This stems from the fact that both negative OCRs and more general types of negative information come from sources independent of the brand. Hence, both tend to be credible (provided they do not come from obviously nefarious sources) and require a strong
10 brand to counter their effect.
Hypothesis 2: Negative OCRs will have a stronger (negative) effect on the products of weak brands than those of strong brands.
We have chosen two similar product categories that differ in terms of maturity in order to examine these relationships across lifecycle stages. As a product category matures and the number of cumulative OCRs increases, there will generally be a corresponding reduction in uncertainty in the category as well as an increase in the credibility of the information contained in the reviews. Whereas a single review may be attributed to the idiosyncratic experiences or motivations of the reviewer, large numbers of consistent reviews will be more reliable. In addition, consumers become more knowledgeable posters and consumers of reviews. Thus, the credibility and impact of the information contained in OCRs tends to increase as they
accumulate, and brands themselves become less influential as uncertainty in the category decreases. Hence, it is possible that strong brands will not be as resistant to the influence of OCRs in more mature categories.
However, as categories mature, consumers gain additional knowledge about the performance and quality of the brands within the category. This brand equity is in large part category-specific since most brands are stronger in some categories than others. For example, Apple is stronger in smart phones than personal computers and Dodge is stronger in trucks than cars. In an emerging category like Blu-ray players, initial brand equity will be based on higher-level categories (e.g., consumer electronics) or related product categories (e.g., DVD players) and will be somewhat uncertain. Over time, it will become more concrete and based more on the focal category. Since brands serve as repositories of product information, the category-specific equity associated with successful brands should become stronger than earlier in the lifecycle as
11
consumers learn more about the performance of the brand in the category. Thus, the relative advantage of strong brands will tend to increase over time, balancing out any potential increase in the credibility of OCRs. This leads us to hypothesize that strong brands will maintain their resistance to the influence of OCRs (either positive or negative) in maturity and that OCRs will impact strong brands no more at maturity than earlier in the lifecycle. Hence, Hypotheses 1 and 2 will generalize across categories.
Hypothesis 3: The moderating effects of brand equity identified above will generalize across both the emerging and mature product categories.
DATA, MODELS, AND ESTIMATION Data
The Blu-ray player category was selected because it was emerging at the time of data collection and appeared to have a variety of strong and weak brands, including those extended from closely related product categories. DVD players were selected as a closely matched mature category with similar numbers of models and brands. Data were collected from the Amazon.com site, with the exception of advertising data, which were purchased from the Nielsen Company for the same time period. Data collection started shortly after Amazon began selling Blu-ray players. Sales rank, OCRs, price, and other data for all models in the Blu-ray player category were collected weekly for 47 weeks from November 01, 2008 to September 21, 2009. This sample consists of 2,324 observations in an unbalanced panel structure of 78 individual models and 47 periods. A total of 3,341 OCRs were observed; 791 were posted in or before the first week and an average of 55.4 were posted in each of the following weeks. While there was considerable fluctuation, the number of additional OCRs decreased by an average of .73 per week. The product selection differed from week to week because some models were introduced or
12
discontinued during the data collection period. Used, refurbished and bundled models were excluded. Of the observations in the sample, 27.19% were models offered by Amazon, the rest were listed on Amazon but sold by other merchants.
Data were collected from the DVD player category from November 01, 2008 to June 06, 2009. This sample consists of 1,080 observations in an unbalanced panel structure of 51 models and 32 weeks. A total of 1,664 OCRs were observed; 971 posts were made in or before the first week and an average of 22.4 were added in each of the following weeks. However, 11 models were added (or returned) to Amazon’s site during the second week, and they accounted for 324 of the 328 reviews that week. After the first two weeks, an average of 13.5 OCRs were posted each week and there was no significant time trend.
In both categories, price includes list price plus shipping and handling costs. If a model was sold by more than one merchant, we used the lowest price charged for a new model. Following the practice of Amazon, and noting that a 3 star rating is below the mean of our data (3.88 for Blu-ray and 3.62 for DVD players), we classified 3 star reviews as negative. Therefore, customer review measures include the number of positive reviews (4 or 5 stars) and the number of negative reviews (1, 2 or 3 stars). We check the robustness of this classification below.
Even though sales data are not accessible, (current) sales ranks are displayed on Amazon for both categories. Hence we use the inverse sales rank for each model as an indicator of sales response. Previous research has found that for many product categories, the relationship between sales rank and sales can be described by a Pareto distribution (i.e., the 80/20 rule), which means that the relationship between ln(sales) and ln(sales rank) is approximately linear; i.e., ln(sales) ≈ a + b * ln(sales rank). This linear relationship has been found to hold for products like books, software, yogurt, women’s clothing, and electronic products (e.g., Brynjolfsson et al. 2003;
13
Brynjolfsson et al. 2011; Ghose and Sundararajan 2006; Goolsbee and Chevalier 2002;
Rosenthal 2005; SymphonyIRI! Group 2011). While we did not run a purchasing experiment to verify this assumption in the two categories, the concentration of products by brand
(HiDefForum 2011) and the concentration of OCRs in the data would suggest it is reasonable. Assuming the Pareto relationship holds approximately, the only differences between a linear model using ln(sales) and one using ln(sales rank) are that the estimated coefficients and their standard errors are scaled by a constant and the estimated intercept is shifted by another constant. Neither of these changes the signs or significance of our coefficients. The negative of ln(sales rank), or ln(1/sales rank), is used in subsequent equations to make the signs of the coefficients easier to interpret (i.e., positive coefficients indicate a greater sales response).2
The Nielsen Company provided weekly advertising expenditures for brands in the Blu-ray and DVD player categories from October 1, 2008 to October 3, 2009. Paid advertisements were placed in newspapers, magazines, TV, radio, and on the Internet. More than 70% of the total $10M in advertising spending was made by Samsung and Toshiba. LG was the only other company to spend more than $500K. Descriptive statistics are presented in Table 1.
--- Insert Table 1 about here --- Models and Estimation
We estimate a three equation model in which brand strength is allowed to vary over time and sales and OCRs are endogenous. The first step is to classify brands into strong and weak
2Use of ordinal scales with more than 4 values (e.g., 5-point Likert scales) as interval data in regressions does not seem to affect type I and type II errors dramatically and is the norm in contemporary social science. The numbers of ordinal values are 72 and 46 in the Blu-ray data and DVD data, respectively. Estimating ordinal regressions with 71 and 45 logit (or probit) functions is impractical. There is a precedent in that a large number of papers in major journals have analyzed Amazon rank data using regression techniques (e.g., Archak, Ghose, and Ipeirotis 2001; Brynjolfsson, Hu, and Smith 2003; Chevalier and Goolsbee 2003; Chevalier and Mayzlin 2006; Ghose and Sundararajan 2006; Sun 2012).
14
categories based on their brand equity. Following Sriram, Chintagunta, and Neelamegham (2006) and Sriram, Balachander, and Kalwani (2007), we use sales data on individual models to
classify brands as strong or weak in the category. In this regression, the brand-dummy
coefficients capture the additional sales impact after the other factors that might influence sales (e.g., advertising, own and competitive OCRs and prices, the total number of models offered, model specific effects, etc.) have been accounted for. Importantly, they capture the impact of a brand on all models in its product line. We use a dynamic specification in which the brand equity regression is re-estimated weekly in order to allow weak brands to grow into strong ones and for formerly strong brands to fall from this category. This is consistent with past research (e.g., Horsky, Misra, and Nelson 2006; Kamakura and Russell 1993).
Specifically, in a model-level regression, we regress ln(Rit) on a series of individual brand dummies and control variables as follows:
(1) ln(Rit) = β0 +
J j j 1 ln(Ri,t-j) + βCUCUi … + βYAYAi + βcPosln(cPosit) + βcNegln(cNegit)
+ βcPos_nln(cPos_nit) + βcNeg_nln(cNeg_nit) + βPln(Pit) + βP_nln(P_nit) + βAdvln(Advit) + βNln(Nt) + βAAit + µi + it , where,
Rit is (1/sales rank) of model i in period t.
CUi …YAi are brand dummies indicating the brand of model i. Brands are listed in Table 2. cPosit (cNegit) is the cumulative number of positive (negative) OCRs for model i in period t. cPos_nit (cNeg_nit) is the total cumulative number of positive (negative) OCRs for all other
models in period t.
Pit (P_nit ) is the price of model i (the average price of all other models) in period t.
15
Nt is the total number of models offered at Amazon in period t.
Ait is a dummy indicating whether model i was offered by Amazon (Ait = 1) or by another
merchant (Ait = 0) at time t.
µi is the time-invariant model-specific effect which captures differences across the models of a brand, such as quality and features like Internet and Wi-Fi capability, etc., and it is an idiosyncratic error.
--- Insert Table 2 about here ---
It is possible that unobserved (by the researcher) product characteristics (such as product quality) influence both sales and OCRs, so ln(cPosit) and ln(cNegit) may be correlated with the model-specific effect µi. In addition, a shock in sales rank for a model may lead to a change in the cumulative number of OCRs, so ln(cPosit) and ln(cNegit) may be correlated with it.
These two possible endogeneity problems prevent the use of random effects estimation of Equation 1 which requires the assumption that all explanatory variables are strictly exogenous with respect to the individual effects (Mundlak 1978). Additionally, fixed effects estimation removes all time invariant effects making it impossible to estimate the brand equities. Therefore, we use an approach suggested by Hausman and Taylor (1981). Applying this method, we use time demeaned (i.e., mean centered within model) values of ln(cPosi,t-1) and ln(cNegi,t-1) as instruments to ln(cPosit) and ln(cNegit) to estimate equation (1) using a random effects estimation method. The time demeaned values of ln(cPosi,t-1) and ln(cNegi,t-1) are valid instruments because they are orthogonal to both the model-specific effect, µi, and the
idiosyncratic error, it, while being correlated with the associated endogenous variables. Once we have determined the strong brands, we use the following model-level equation
16
to estimate the main and interaction effects of OCRs and brand equity on model sales rank:
(2) ln(Rit) = β0 +
J j j 1 ln(Ri,t-j) + βcPosln(cPosit) + βcNegln(cNegit) + βBBit + βBcPosBit*ln(cPosit)
+ βBcNegBit*ln(cNegit) + βcPos_nln(cPos_nit) + βcNeg_nln(cNeg_nit) + βAdvln(Advit) + βPln(Pit) + βP_nln(P_nit) + βNln(Nt) + βAAit + µi + it
where, in addition to the previously defined variables,
Bit is a strong-brand dummy (Bit = 1 if the brand of model i is significantly positive at period t which is estimated from Equation 1 using the data in the first t-1 periods, Bit = 0 otherwise).
As in Equation 1, ln(cPosit) and ln(cNegit) may be correlated with both µi and it so we cannot use random effects methods. However, unlike Equation 1, we are not interested in time invariant variables in Equation 2 so we can use an estimation method suggested by Arellano and Bond (1991) which allows us to utilize more instruments. The first step is to first-difference the model to eliminate all of the model-specific effects, µi. Equation 2 becomes:
(2’) Δln(Rit) =
J j j 1 Δln(Ri,t-j) + βcPosΔln(cPosit) + βcNegΔln(cNegit) + βBΔBit
+ βBcPosΔ[Bit*ln(cPosit)] + βBcNegΔ[Bit*ln(cNegit)] + βcPos_nΔln(cPos_nit) + βcNeg_nΔln(cNeg_nit) + βAdvΔln(Advit) + βPΔln(Pit) + βP_nΔln(P_nit) + βNΔln(Nt) + βAΔAit + Δit
Then we use lags of ln(cPosit), ln(cNegit), Bit*ln(cPosit), and Bit*ln(cNegit), up to t-2 and
lags of ln(Ri,t-j) up to t-j-1 as instruments for their first-differences, Δln(cPosit), Δln(cNegit),
Δ[Bit*ln(cPosit)], Δ[Bit*ln(cNegit)], and Δln(Ri,t-j), respectively, to perform a GMM estimation of Equation 2’. These lags are valid instruments because they are uncorrelated with Δit while correlated with the first-differences of the endogenous variables.
17
Equations 2 and 2’ so the first-differencing can be ignored when interpreting the coefficients. While both first-differencing and a fixed effects transformation can eliminate the model-specific effects µi, first-differencing is used because it involves only data in periods t and t-1 so we can
use all data from period 1 up to period t-2 as instruments. A fixed effects transformation uses the
data from all time periods so it renders all lags useless as instruments (see Nickell 1981 and Roodman 2006 for the mathematics). Specifically, a fixed effect transformation of ln(Rit) (i.e., mean centering within each model) involves the value of ln(Rit) in all periods (in the model mean calculation) which makes the error term correlated with all the lags of ln(cPosit) and ln(cNegit).
Because it has been found that an increase in sales will lead to the posting of more OCRs (Duan, Gu, and Whinston 2008a, 2008b), we use Equations 3 and 4 to examine the effects of sales rank on the number of positive and negative OCRs as follows:
(3) ln(Posit) = α0 +
L l l 1
ln(Posi,t-l) + αRln(Rit) + αBBit + αdln(dit) + αd2[ln(dit)]2 + i + it(4) ln(Negit) = 0 +
M m m 1
ln(Negi,t-m) + Rln(Rit) + BBit + dln(dit) + d2[ln(dit)]2 + i + itwhere, in addition to the previously defined variables,
Posit (Negit) is the number of positive (negative) OCRs generated for model i in period t. dit is the listed duration of model i at Amazon until period t.
i and i are time-invariant model-specific effects and it and it are idiosyncratic errors. As above, it is possible that unobserved product characteristics influence both sales and OCRs, so ln(Rit) may be correlated with the model-specific effects i and i. In addition, a shock in the number of positive or negative OCRs may lead to a change in sales rank, so ln(Rit) may be correlated with it and it. Therefore, we use the same method used to estimate Equation 2 to
18
estimate Equations 3 and 4. The differences are that we use lags of ln(Rit) up to t-2 and lags of
ln(Posi,t-l) and ln(Negi,t-m) up to t-l-1 and t-m-1 as instruments for their first-differences, Δln(Rit),
Δln(Posi,t-l), and Δln(Negi,t-m). All equations are estimated separately in the Blu-ray and DVD player categories below.
RESULTS Brand Classification and Dynamics
In both categories, Equation 1 was first estimated with Samsung, the highest ranked brand in Best Global Brands 2009 (Interbrand 2009) and Technology Brands Top Ranking 2009 (BrandZ 2009) (see Table 2), as the base, or reference, brand. Those brands which had
significant negative intercepts were then all set as the base, or reference, brands to re-estimate Equation 1. Brands which had significant positive intercepts in the re-estimated model were coded as strong brands. The first 14 weeks were used as a calibration period. Recognizing that brand strength is dynamic, we updated these brand classifications weekly based on t-1 weeks of
data (e.g., we use data from week 1 to week 14 to estimate brand strength in week 15). In the Blu-ray category, Sony and Samsung, were classified as strong brands for the entire period. Panasonic and LG were initially classified as weak brands, but each grew its brand equity within the category over time. Panasonic became a strong brand in week 15 and LG in week 26. Oppo did not enter the market until week 38, but became a strong brand in week 40. None of the strong brands reverted to being a weak brand over time.
In the more mature nature of the DVD category, LG, Oppo, Panasonic, Philips, Pioneer, Samsung, and Toshiba were strong brands from the start of our data window. Sony was not initially a strong brand. However, after launching several new models in weeks 22 and 23, it achieved strong brand status in week 28. None of the strong brands reverted to being a weak
19 brand over time.
The Effect of OCRs on Brand Equity
Because several brands made the transition from being weak to strong brands, we looked at the impact that cumulative OCRs had on this transition. Because only one brand made the transition in the DVD player category, we restrict our attention to the Blu-ray category.
Figure 1 shows the cumulative number of positive OCRs generated for the leading model of each of the ten brands of Blu-ray players that started as weak brands and had one or more models on the market for our entire data collection period. Clearly, models of Panasonic and LG, which became strong brands, generated many more positive OCRs (and proportionally fewer negative OCRs) than their competitors. The flattening of the curves for these two brands
indicates that most of the OCRs were generated within the first 20 to 30 weeks. Figure 2 expands this to show the total cumulative number of positive OCRs for each brand across all models. Here, the curves for these two brands continue to rise, indicating that as new Panasonic and LG models were introduced, they also generated positive OCRs. Finally, a comparison of the two figures shows that most of LG’s OCRs in the first twenty weeks were generated by one model, but more than one Panasonic model was generating a substantial number of OCRs.
--- Insert Figures 1 and 2 about here ---
In order to see how Panasonic and LG differed from the other eight brands, we estimated a proportional hazard function on the probability that brand k became a strong brand in period t,
given that it was not a strong brand previously, at the brand-level on these ten brands:
(5) hk(t, Xk,t) = h0(t) Φ , where:
h0(t) is the baseline hazard rate.
20
Φ is a vector of estimated coefficients.
k is the time-invariant brand-specific effect which represents the combined effect of all omitted brand-specific covariates.
Because only two brands became strong during this period, we focus on a parsimonious model with four independent variables: cumulative numbers of positive and negative OCRs for the brand, advertising expenditures, and the number of models in the product line of the brand. Once again, we face potential complexities in that the cumulative numbers of positive and negative OCRs can be endogenous in two ways. First, a change in brand strength may lead to a change in OCRs. Second, OCRs might be correlated with k. We address the first issue by using lags of the cumulative numbers of positive and negative OCRs to instrument for them. For the second issue, estimations of Equation 5 with and without k yielded the same results, indicating that we could eliminate k so the correlation between k and OCRs is not a concern.
The results indicate that making the transition to a strong brand is highly related to the cumulative number of positive OCRs, but is not related to the cumulative number of negative OCRs, advertising expenditures, or the number of models sold by the brand (see Table 3). We also estimated Equation 5 based on the cumulative number of positive and negative OCRs for just the leading model of each brand (which is not confounded by the number of models the brand sells). This analysis also found that brand strength was significantly related to the
cumulative number of positive OCRs but not to negative OCRs, advertising expenditures, or the length of the product line.
21 The Moderating Effect of Brand Equity
After determining the strong brands each week, we estimated Equations 2, 3, and 4. We determined the appropriate number of lags based on the method suggested by Arellano and Bond (1991). In both categories, tests (Table 4) indicated that the instruments are valid and the
estimations do not suffer from serial correlation problems.3 --- Insert Table 4 about here --- Blu-ray Players
Table 5a presents results for Blu-ray players without and with the brand interactions. The first column in Table 5a shows the elasticities of cumulative positive and negative OCRs on sales rank for all models (i.e., both strong and weak brands). The sales rank elasticity with respect to the cumulative number of positive OCRs (.568) is significant, but the sales rank elasticity with respect to the cumulative number of negative OCRs (-.327) is not significant.
In the second column of Table 5a, the elasticities of cumulative positive and negative OCRs for the models of weak brands are given by the main effects. The cumulative number of positive and negative OCRs for the models of weak brands have significant elasticities (1.091 and -.579) on their sales ranks. The interactions indicate significant differences between the elasticities for the models of strong and weak brands for both cumulative positive and negative OCRs (-1.154 and .902). The elasticities of cumulative positive and negative OCRs for the models of strong brands are the sums of the main effects and interactions (-.063 and .323);
3 The Arellano-Bond (1991) tests for second-order serial correlation in the first differences Δ
it, Δit, and Δit are insignificant, indicating no evidence of first-order serial correlation in it, it, and it. The Sargan (1958) tests for over-identifying restrictions are insignificant, so we cannot reject the joint validity of the instrument sets used to estimate Equations 2, 3, and 4. The difference-in-Sargan tests of exogeneity of the instrument subsets used to estimate the three equations are also insignificant, indicating that all instrument subsets are exogenous.
22
neither of these elasticities are significant (p-values of .870 and .367, respectively).4 However, the models of strong brands receive a substantial sales boost from brand equity as seen by the sales multiplier of 2.186 associated with the strong brand dummy. Thus, our results show that the models of weak brands are affected more by both cumulative positive and negative reviews, supporting Hypotheses 1 and 2 in this category.
--- Insert Table 5 about here ---
The effects of the other variables are similar in the first and second columns and of the expected signs with or without interactions. The price of model i and the cumulative number of
positive reviews for other models are significantly and negatively related to the sales rank of model i. The total number of models offered at Amazon has a marginal negative effect on the
sales rank of model i. The price of other models is significantly and positively related to the sales
rank of model i. However, being offered by Amazon, advertising expenditures for the brand, and
the cumulative number of negative reviews for other models are not significantly related to the sales of model i.
DVD Players
Table 5b presents results for the DVD player category. In contrast to the Blu-ray analysis, neither cumulative positive nor negative OCRs are significantly related to sales rank in the first column. Hence, OCRs appear to have less effect, without considering strong and weak brands.
When the interactions with brand strength are included in the second column of Table 5b, the cumulative number of positive and negative OCRs for the models of weak brands again have significant elasticities (1.192 and -.972) on their sales ranks. There are significant interactions
4 To calculate significance, we reversed the coding of the brand dummy variable from identifying strong brands to
identifying weak brands. Under this formulation, the impact of cumulative OCRs on models of strong brands is given by the main effect coefficients. Neither of these was significant.
23
between brand strength and cumulative positive OCRs (-1.274) as well as cumulative negative OCRs (1.018).5 Summing the main effects and interactions, we see that neither cumulative positive nor negative OCRs have a significant elasticity with the sales ranks of models of strong brands (p-values of .545 and .704, respectively). However, again, the sales multiplier associated with the strong brand dummy variable of .718 indicates a substantial benefit to models of these brands. Hence, support is also found for Hypotheses 1 and 2 in this category and thus the generalizability of the results per Hypothesis 3. OCRs appear to matter about equally and brand equity exhibits moderating effects of approximately the same magnitude.
In the first column of Table 5b, the only other variable significantly related to sales rank (in addition to lagged sales rank) is own model price. In the second column, the number of cumulative positive OCRs for other models is also significantly and negatively related to sales of the focal model whereas cumulative negative OCRs for other models are insignificant.
Advertising expenditures are controlled, but are insignificant. This may be explained in part by the relatively infrequent nature of category-specific advertising observed in both categories. The Role of Brand Equity and Sales in OCR Generation
Turning to the factors associated with the generation of positive and negative OCRs, estimates for Equations 3 and 4 appear in Table 6 for both Blu-ray and DVD players. Sales rank significantly and positively predicts the number of positive OCRs, but the effect of sales rank on
the number of negative OCRs is insignificant. In addition, the effects of listed duration on both the number of positive and negative OCRs are positive but the effects of its quadratic term are
5 The p-value of .053 is marginally significant. However, we are using two-tailed confidence intervals despite the
a priori directional hypotheses. A formally correct one-tailed test is significant at p ≤ .05. Moreover, we produce a lower p-value in our robustness check below where 4 and 5 star reviews are counted as positive and only 1 and 2 star reviews are counted as negative (p = .043, two-tail), giving confidence in the result.
24
negative. This means that the longer a model has been listed on Amazon, the more (both positive and negative) OCRs are posted for it, but after a peak (in all regressions, this peak occurs less than three weeks after introduction), fewer OCRs are posted for it. Hence, both positive and negative reviews increase and then decrease over time, but greater sales tend to generate only additional positive reviews. Thus, there is no penalty in terms of more negative reviews being generated for more popular models. Finally, the models of strong brands do not generate more OCRs, either positive or negative, after the effect of sales is controlled. Models of strong brands might be expected to attract more attention and generate greater customer involvement, both of which would result in more OCRs, but this is not the case.
--- Insert Table 6 about here --- Robustness Check
Our main analysis is based on the number of positive and negative OCRs, where 4 and 5 star reviews are classified as positive and 1, 2, and 3 star reviews are classified as negative following Amazon’s practice. Because of questions about how 3 star reviews should be
classified, we keep the same definition of positive reviews and classify reviews with 1 or 2 stars as negative to check the robustness of the analysis to this classification.
The first column of Table 7 repeats the second column of Table 5 for ease of comparison. The second column contains the coefficients when all of the OCRs with three stars have been dropped. There is no meaningful change from the first column in terms of control variables so they are not presented in Table 7. The primary change is that all of the cumulative OCR and interaction coefficients (except the cumulative positive OCR coefficient for Blu-ray players) are larger in absolute value owing to the greater difference between positive and negative OCRs. The interpretations are quite similar; hence, our results seem to be robust to a different
25 classification of positive and negative reviews.
--- Insert Table 7 about here --- Comparison to Volume and Valence Approach
Previous studies have found a significant positive relationship between the number of OCRs and sales (e.g., Chen, Wu, and Yoon 2004; Chevalier and Mayzlin 2006; Dellarocas, Zhang, and Awad 2007; Li and Hitt 2008; Liu 2006). However, the relationship between the valence of OCRs (average customer rating) and sales is mixed. While some studies have shown that the valence of customer reviews affects sales positively (e.g., Chevalier and Mayzlin 2006; Dellarocas, Zhang, and Awad 2007; Karniouchina 2011; Li and Hitt 2008), others have found an insignificant relationship (e.g. Chen, Wu, and Yoon 2004; Duan, Gu, and Whinston 2008a, 2008b; Liu 2006). Duan, Gu, and Whinston attribute the result to controlling endogeneity. If one takes the position that positive reviews help sales while negative reviews hurt, it may be better to estimate these two effects separately than estimating the effect of their sum and average. Hence, we compare our approach with the volume and valence approach.
The third column of Table 7 replaces cumulative positive and negative OCRs with cumulative total OCRs and valance. First, as above, most of the control variables retain the same size, sign, and significance as those in the first column so they are not presented in Table 7. Turning to the key comparison, in column 3, the cumulative total number of OCRs is significantly related to sales but the coefficient is smaller than the coefficient of cumulative positive OCRs in the first column. Hence, it is possible that the effect of cumulative total OCRs is a compromise between the positive effect of cumulative positive OCRs and the negative effect of cumulative negative OCRs. Note that both Blu-ray and DVD players have many more
26
The coefficient of the interaction between cumulative total OCRs and brand strength is significantly negative, indicating that the products of strong brands benefit less from cumulative total OCRs, which is consistent with our theory. An estimation with reverse brand coding shows that cumulative total OCRs have no significant effect on sales of the models of strong brands (p-values are .380 and .131 for Blu-ray and DVD players, respectively). Meanwhile, the
coefficients of both valence and its interaction with brand strength are insignificant. Thus, we see a pattern where only volume (not valence) matters as has been reported in the studies mentioned above. Since the volume and valence approach provides fewer insights than ours, it may be better for both researchers and managers to account for positive OCRs and negative OCRs separately.
DISCUSSION Summary of Findings and Theoretical Implications
In this study we provide evidence across two product categories that brand equity moderates the relationship between OCRs and sales. In both categories, cumulative positive OCRs increase, and cumulative negative OCRs decrease, the sales of models of weak brands. In contrast, neither cumulative positive nor negative OCRs have a significant effect on the sales of models of strong brands. However, these models do receive a significant sales boost from being part of a strong brand; hence, they are not disadvantaged due to their resistance to positive OCRs and are protected to a degree from negative OCRs. In the reverse direction, greater sales lead to more positive OCRs, but are not significantly related to the number of negative OCRs. We find a decline in the number of positive and negative OCRs that customers post over time, although the relationship is curvilinear in that OCRs increase for the first three weeks, and then decrease.
27
The results of the hazard model for the Blu-ray category show that cumulative positive OCRs help build the equity of weak brands, whether the OCRs pertain to all models in the product line or only the leading model. Interestingly, cumulative negative reviews do not significantly influence this transition in brand strength, indicating that their effects are largely constrained to the individual model itself. Hence, weak brands do not seem to be held back by cumulative negative reviews so much as they are helped by cumulative positive reviews.
Combined with the fact that sales are shown to generate more positive than negative OCRs, this creates a positive feedback loop for the models of a weak brand in which sales lead to positive
OCRs and greater brand equity, which loops back to positively affect sales for all of the models of that brand.
In contrast to the speculation that brands will matter less as OCRs become more readily available, we find that OCRs matter less for strong brands. In essence, brand equity tends to trump OCRs more than the other way around. Much like the importance of brands in the face of stronger more concentrated retailers and their private labels (the last purported “brand-killer”), brands will continue to be a critical factor for companies competing in an online environment. The results also suggest that corporate brand equity may not be a good indicator of strength in a specific category. For example, Philips has the third largest corporate brand value in the
Interbrand ranking after only Samsung and Sony, but in the Blu-ray disc player category its brand equity is less than that of LG, Panasonic, and Oppo.
The results across both categories show a pattern in which negative reviews tend to be less impactful than positive reviews; viz., the elasticities for cumulative negative reviews are lower than for cumulative positive reviews in both categories and cumulative negative OCRs for other products are not significantly related to sales of the focal product whereas cumulative
28
positive reviews for other products are. The lesser effects associated with negative reviews are interesting because they go against perspectives such as prospect theory in which losses loom larger than gains, which should cause consumers to put more weight on negative reviews as they indicate downside risk. One explanation is that there are more positive reviews than negative reviews. It is possible that this reduces the credibility and influence of the latter.
The effects of product category maturity are complex. On one hand, the regressions presented in Tables 5 and 6 indicate very similar relationships among OCRs, sales and brand strength across the emerging and mature categories. On the other hand, we see a much smaller number of OCRs posted each week, a larger number of strong brands, and less movement from weak to strong brands in the mature category. Together, these observations suggest that as a product category matures, there may be less room for new strong brands, and that if a brand has not made the move from weak to strong in the first few years of a product category, the chances that it will do so in the future are smaller. However, even in more mature categories, the results show that it is possible for models of new or weak brands to use OCRs to increase sales. Furthermore, in many mature categories, current leading brands are not pioneers, or even fast seconds, but brands that entered the category around ten years after the pioneer (Golder and Tellis 1993). While not all later entrants are going to become leaders (or strong brands), it does appear that OCRs provide a way for them to accomplish this task.
Finally, the findings suggest that positive OCRs are different from producer-generated marketing communications, the effectiveness of which is boosted by increased brand equity. We argued above that the greater credibility of OCRs implies that they have less need of a strong brand to lend credibility like advertising does. In addition, typical marketing communications are passively received, giving those for more familiar brands with established cognitive schemas a
29
better chance of being attended to and remembered. In contrast, OCRs are processed more actively since they must be sought out and read. This active processing may undo the bias in favor of strong brands, especially if the brand strength gives consumers a feeling of confidence about the performance of strong brands so that they attend more to reviews for weaker brands. Managerial Implications
Most previous studies involving online word-of-mouth (eWOM) have found a positive relationship between eWOM and sales. However, this body of research has not looked at
traditionally branded products. It has implicitly advocated strategies to increase the generation of eWOM for all products (an exception is Zhu and Zhang (2010), as discussed above). In contrast, our findings suggest very different strategies for the models of strong and weak brands.
Models of weak brands should focus on generating positive OCRs since they benefit sales of that model directly as well as the equity of the brand as a whole. First, in the pursuit of sales, individual models of weak brands benefit from a sizable cumulative positive OCR elasticity. Second, they benefit from the positive feedback loop between sales and positive OCRs. Finally, the generation of a large number of positive OCRs for one or more models is strongly associated with increased brand equity, which benefits all models in the line, not just the one with increased positive OCRs.
These paths give weaker brands a way to compete other than through traditional
marketing communications, which favors strong brands. Especially for weak brands, the results support currently popular ideas such as “flipping the funnel” where marketing dollars are removed from mass media advertising and focused more on increasing satisfaction, retention, and positive word-of-mouth which is then used to drive the acquisition process.
30
that capture consumer excitement. Second, getting online distribution may be easier than off-line distribution due to the larger number of models carried by an online distributor like Amazon in comparison to even a very large bricks and mortar store (Brynjolfsson et al. 2003). An additional benefit is that online distribution usually comes with a vehicle for posting reviews. Furthermore, even though both online and off-line distributors tend to have Pareto-like concentrations among more popular products, the online channel tends to have a longer tail (Brynjolfsson et al. 2011), which makes it easier for models of less well-known brands to capture some initial sales. Then, they can generate more positive OCRs by providing seeds for and facilitating positive OCR generation through actions such as:
Making detailed information about products available and easily accessible. There is anecdotal evidence that customers refer to information from producers in their reviews.
Establishing brand communities and early adopter clubs. Members of these clubs can buy products with incentives before launch to spark the feedback process. Producers can use positive feedback as seeds and negative feedback to modify their products before launch.
Providing samples to expert review websites. There is again anecdotal evidence that customers refer to expert reviews in their own reviews.
Sending reminders and incentives to customers to encourage the posting of reviews. In contrast to weak brands, additional positive OCRs do not further benefit the models strong brands. Hence, strong brands should consider actions to build brand equity through advertising and promotions rather than relying too much on OCRs. As we caution above, brand equity tends to be category-specific, so managers should understand the equity of their brand within a given category, especially when new. While models of a strong brand are not affected by their own OCRs, they are affected by positive OCRs for models of weak brands as consumers
31
are taken away from them. Hence, competing with weak brands in the OCR space may not be a winning strategy for strong brands. They may be better off investing in other marketing
practices.
A closely related implication is that the role of OCRs is likely to change as the brand evolves over time. Newer, weaker brands should focus on OCRs and the synergistic feedback
loop, promoting positive and trying to minimize negative feedback. However if these brands are able to increase brand equity, additional OCRs become less impactful; hence a shift in strategy toward a more balanced approach to marketing and maintaining brand strength may be required. While controlling negative OCRs may not be as important for stronger brands, they are still probably very well advised to monitor negative OCRs and take corrective actions. At minimum, the cumulative body of negative OCRs will matter if these brands lose their strength over time.
In contrast to the results of Zhu and Zhang (2010), the results of this study suggest that OCRs are important for models of weak brands even before they are launched and certainly during the first several months of launch. Additionally, rather than just being a tool for niche products, OCRs represent a way for a model to help build overall brand equity and benefit all (and future) models under the brand umbrella. This occurs because positive OCRs spillover to the brand itself. For managers of traditionally branded products, attending to existing brand equity is an important part of understanding the influence of OCRs. Brand and product
extensions as well as model replacement are situations in which pre-existing brand equity will impact consumer processing of OCRs. Strong brands can rely more on traditional marketing strategies for influencing awareness and trial, whereas weaker brands should focus more on facilitating OCRs as they can have a larger impact on the models of these brands.
32
short-term sales of individual models when anticipating the role of OCRs. According to the perspective developed in this research, strong selling products offered by weak brands may still benefit from positive OCRs, especially if the reason for their popularity is a low price or heavy advertising and promotions which create a disconnect between sales and the inherent appeal of the brand. Indeed, investing in OCRs for these models may be an especially good strategy
because they are already popular despite the weak brand. This may indicate that the product itself is a winner that can be improved by facilitating credible reassurance in the form of OCRs that is not provided by the brand itself. Positive OCRs generated for the model (which are a function of sales) can then further increase the equity of the brand. Models which are popular because of price promotions may also find that positive OCRs allow higher prices and margins.
Finally, one potential implication of the finding that positive reviews for competing products have a negative effect on the sales of a focal product but negative reviews for other products do not is that subterfuge to post negative OCRs as a competitive tactic may not work very well, and worse, may trigger a wave of positive OCRs for the competitive product from customers who tend to argue for their committed brands (Ahluwalia, Burnkrant, and Unnava 2000). This could end up decreasing sales for the product responsible for the deception. This is especially likely if weak brands try to use this tactic against strong brands.
Methodological Implications and Limitations
In this study, we find that attending to both positive and negative reviews is important. We find the same pattern reported in some earlier studies when we use a volume and valence approach where valence does not seem to matter. The more fine-grained examination in this study shows that negative reviews hurt. It is not simply the buzz factor that matters.
33
One explanation for the difficulty encountered in previous volume and valence research is that positive OCRs are generated faster than negative OCRs for products as their sales rank improves, creating the feedback loop for positive OCRs observed in this study. This increases the correlation between the total number and the valence of OCRs as the number of OCRs grows. Coding schemes based on volume and valence may have difficulty teasing this apart. In addition, the effect of cumulative total OCRs on sales may be a compromise between the positive effect of cumulative positive OCRs and the negative effect of cumulative negative OCRs. Thus, we suggest using the number of positive and negative OCRs as distinct variables to accurately reflect consumer generated sentiment.
Moreover, because the effects of OCRs on the sales of models of strong brands are different from those of weak brands, excluding the interactions between brand equity and customer reviews from the equation would bias the results. Therefore, future studies
investigating the effects of OCRs on sales in branded categories should consider this factor. Our treatment of endogeneity is more complete than in most of the extant literature, where only a few papers address this issue. For example, the difference-in-difference approach in Chevalier and Mayzlin (2006) and Zhu and Zhang (2010), as well as the fixed effects estimation method in Zhang and Dellarocas (2006), purge the correlation between cumulative OCRs and the individual product specific effects caused by unobserved factors. The method used in this
research also remedies the correlation between cumulative OCRs and the idiosyncratic error in Equation 2. Without that extra treatment, the results of Equation 2 would still be biased. This suggests that both of the endogeneity problems should be addressed in similar studies.
A recent study by Sun (2012) suggests that OCR variance can impact sales when the average rating is low. While we do not investigate this issue, it is interesting to note that our
34
results appear to be consistent with this finding. A model of a weak brand that received an OCR rating of three from each person would be forecast to have lower sales than a model that received half fives and half ones.
Our study suffers from a number of limitations. First, we have not accounted for the effects of expert reviews and customer reviews on other sites. It is possible that expert reviews (e.g., on CNET) or OCRs on other sites (e.g., Best Buy) also impact sales rank on Amazon. However, the effects of expert reviews are likely to be cancelled out in the first-differencing step of our estimation because experts usually post their reviews when a product is launched and are not likely to change their reviews over time so they are time-invariant. An extension to this research would be to examine the impact of OCRs on other sites.
Second, by estimating brand equity from sales data we introduce measurement error into the second-stage models (Equations 2-4). We believe the dynamic and category-specific nature of the estimates more than makes up for any measurement error in comparison to using
secondary brand rankings, which are themselves subject to measurement error.
Third, advertising is not particularly widespread across brands in these categories. Hence, while we see limited effects of advertising on both sales and brand equity, this may not hold in other categories. The low advertising by smaller brands may suggest that they see the limitations of this strategy and are already relying to a degree on OCRs to support their models and brands.
35 REFERENCES
Ahluwalia, Rohini, Robert E. Burnkrant, and H. Rao Unnava (2000), “Consumer Response to Negative Publicity: The Moderating Role of Commitment,” Journal of Marketing Research, 37
(May), 203–214.
Ailawadi, Kusum L., Donald R. Lehmann, and Scott A. Neslin (2003), “Revenue Premium as an Outcome Measure of Brand Equity,” Journal of Marketing, 67, 1–17.
Albers, Sönke, Murali K. Mantrala, and Shrihari Sridhar (2010), “Personal Selling Elasticities: A Meta Analysis,” Journal of marketing Research, 47 (October), 840 – 853.
Archak, Nikolay, Anindya Ghose, Panagiotis G. Ipeirotis (2011), “Deriving the Pricing Power of Product Features by Mining Consumer Reviews,” Management Science, 58(7), 1485-1509.
Arellano, Manuel and Stephen Bond (1991), “Some Tests of Specification for Panel Data: Monte Carlo Evidence and an Application to Employment Equations,” The Review of Economic Studies,
58(2), 277-297.
Belch, George E. (1981), “An Examination of Comparative and Noncomparative Television Commercials: The Effects of Claim Variation and Repetition on Cognitive Response and Message Acceptance,” Journal of Marketing Research, 18 (Aug), 333–349.
Bickart, Barbara and Robert M. Schindler (2001), “Internet Forums as Influential Sources of Consumer Information,” Journal of Interactive Marketing, 15(3), 31-40.
Boulding, William and Amna Kirmani (1993), “A Consumer-Side Experimental Examination of Signaling Theory: Do Consumers Perceive Warranties as Signals of Quality?” Journal of
36
BrandZ (2009), “Technology Brands Top Ranking 2009,” (accessed June 24, 2010), [available at http://www.brandz.com/upload/brandz-report-2009-complete-report(1).pdf].
Brynjolfsson, Erik, Yu Hu, and Duncan Simester (2011), “Goodbye Pareto Principle, Hello Long Tail: The Effect of Search Costs on the Concentration of Product Sales,” Management
Science, 57(8), 1371-1386.
_______, ________, and Michael D. Smith (2003), “Consumer Surplus in the Digital Economy: Estimating the Value of Increased Product Variety,” Management Science, 49 (11), 540 – 546.
Calder, Bobby J. and Brian Sternthal (1980), “Television Commercial Wearout: an Information Processing View,” Journal of Marketing Research, 27 (May), 173–186.
Campbell, Margaret C. and Kevin L. Keller (2003), “Brand Familiarity and Ad Repetition Effects,” Journal of Consumer Research, 30 (2), 292-304.
Chen, Pei-Yu, Shin-yi Wu, and Jungsun Yoon (2004), “The Impact of Online Recommendations and Consumer Feedback on Sales,” in proceedings of the International Conference on
Information Systems, 25, 711–24.
Chen, Stephen (2001), “Assessing the Impact of the Internet on Brands,” Journal of Brand
Management, 8(4-5), 288-295.
Cheong, Hyuk J. and Margaret A. Morrison (2008), “Consumers' Reliance on Product
Information and Recommendations Found in UGC,” Journal of Interactive Advertising, 8(2),
37
Chevalier, Judith A. and Dina Mayzlin (2006), “The Effect of Word of Mouth on Sales: Online Book Reviews,” Journal of Marketing Research, 43 (3), 345-354.
Dawar, Niraj and Madan M. Pillutla (2000), “Impact of Product-Harm Crises on Brand Equity: The Moderating Role of Consumer Expectations,” Journal of Marketing Research, 37 (May),
215–226.
Dellarocas, Chrysanthos, Xiaoquan Zhang, and Neveen F. Awad (2007), “Exploring The Value of Online Product Reviews in Forecasting Sales: The Case of Motion Pictures,” Journal of
Interactive Marketing, 21(4), 23–45.
Duan, Wenjing, Bin Gu, and Andrew B. Whinston (2008a), “Do Online Reviews Matter? – An Empirical Investigation of Panel Data,” Decision Support Systems, 45, 1007-1016.
———, ———, and ——— (2008b), “The Dynamics of Online Word-Of-Mouth and Product Sales - An Empirical Investigation of the Movie Industry,” Journal of Retailing, 84, 233–242.
Dubin, Jeffrey A. (1998), Studies in Consumer Demand: Econometric Methods Applied to
Market Data. Norwell, MA: Kluwer Academic Publishers.
Erdem, Tulin and Joffre Swait (1998), “Brand Equity as a Signaling Phenomenon,” Journal of
Consumer Psychology, 7 (April), 131–57.
Erdem, Tulin, Joffre Swait, and Ana Valenzuela (2006), “Brands as Signals: A Cross-Country Validation Study,” Journal of Marketing, 70 (January), 34–49.
Ghose, Anindya and Arun Sundararajan (2006), “Evaluating Pricing Strategy Using Ecommerce Data: Evidence and Estimation Challenges,” Statistical Science, 27(6), 995 – 1011.
38
Golder, Peter N. and Gerrard J. Tellis (1993), "Pioneer Advantage: Marketing Logic or Marketing Legend?" Journal of Marketing Research, 30 (May), 158-70.
Goolsbee, Austan and Judith A. Chevalier (2002), “Measuring Prices and Price Competition Online: Amazon and Barnes and Noble,” Working Paper No. 02-23, Yale ICF.
Hansen, Lars P. (1982), “Large Sample Properties of Generalized Method of Moments Estimators,” Econometrica, 50, 1029-1054.
Hausman, Jerry A. and William E. Taylor (1981), “Panel Data and Unobservable Individual Effects,” Econometrica, 49 (6), 1377-1398.
HiDefForum (2011), “Next Gen DVD Players (Blu-ray) Outsell Home DVD Players in Latest NPD $ Sales Report,” (accessed November 3, 2011), [available at
http://www.highdefforum.com/high-definition-media/98117-next-gen-dvd-players-blu-ray-outsell-home-dvd-players-latest-npd-sales-report.html].
Hoeffler, Steve and Kevin L. Keller (2003), “The Marketing Advantages of Strong Brands,”
Journal of Brand Management, 10 (6), 421–445.
Horsky, Dan, Sanjog Misra, and Paul Nelson (2006), “Observed and Unobserved Preference Heterogeneity in Brand-Choice Models,” Marketing Science, 25 (4), 322–335.
Hsu, Jane L. and Gary S. Liu (2000), “Consumer Perceptions of Fluid Milk Advertising in Taiwan,” International Journal of Advertising, 19, 471–486.
